I was looking through some stuff online and it struck me that an Intelligent Agent architecture would be fairly easy to implement, using only Open Source/Free Software tools.
The Agent would be running as a daemon, built using Perl and easily available Perl modules, for SOAP, CGI, AJAX and whatever else you need. Early in development, you'd have a collection of stump functions to call, and per Paul Graham, once you have a working Agent, you can add and extend functionality.
On the client side, you'd run Mozilla, and your agent communication code would run as a GreaseMonkey script set to run on every page, and probably sending the URL, REFERER and a few other points for each page.
But what data?
I'm not actually sure yet. Depends on context. Depends on what you want. I can come up with a few usage scenarios, though.
Comment if you're interested, or at least think I'm on the good kind of crack.